{ "schema_version": 1, "builder_plan": { "scenario": "Start with one repeated product task that has a measurable failure or cost.", "change": "Read the methods and identify one change you can isolate in that workflow.", "baseline": "Your current implementation on the same held-out tasks.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Successful outcomes, response time, human rework, and total cost per successful task.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2609.13947", "title": "Hardware-Aware Learned Representation Compression for Distributed In-Sensor Vision", "summary": "OASIS replaces transmission of raw high-resolution sensor images (and conventional DNN partitioning) with compact task-relevant latents produced by a lightweight encoder on a CIS-integrated logic chip, using 4-bit quantization plus Huffman coding or Sobol-based hyperdimensional computing. It costs a custom FPGA/ASIC near-sensor pipeline, end-to-end training with task, entropy, and reconstruction objectives, and up to roughly one percentage point of accuracy at the most aggressive compression. It can fail when the target task drifts from the training distribution (the latent is task-specific), when the required CIS-plus-logic hardware or FPGA toolchain is unavailable, or when associative-memory HDC classification underperforms on complex dense-prediction workloads.", "rationale": "The technique requires custom sensor-integrated silicon, FPGA implementation, and board-level hardware co-design, none of which fits a production practice limited to one 24 GB GPU or third-party APIs.", "technique": "Hardware-aware in-sensor representation compression (OASIS)", "family": "outro", "recommendation": "nao_aplica", "published": "2026-09-12", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.13947", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "OASIS replaces transmission of raw high-resolution sensor images (and conventional DNN partitioning) with compact task-relevant latents produced by a lightweight encoder on a CIS-integrated logic chip, using 4-bit quantization plus Huffman coding or Sobol-based hyperdimensional computing. It costs a custom FPGA/ASIC near-sensor pipeline, end-to-end training with task, entropy, and reconstruction objectives, and up to roughly one percentage point of accuracy at the most aggressive compression. It can fail when the target task drifts from the training distribution (the latent is task-specific), when the required CIS-plus-logic hardware or FPGA toolchain is unavailable, or when associative-memory HDC classification underperforms on complex dense-prediction workloads.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.13947", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "OASIS reduces total system energy by approximately 2x-4.5x across visual wake-word, hand-tracking, and eye-tracking tasks, with up to 18,816x communication reduction versus raw 8-bit image transmission (64-dim hypervector, under 1 percentage point accuracy loss).", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.13947", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [ { "anchor": "S3.E1", "label": "(1)", "section": "III Proposed Method", "role": "proposed_method", "latex": "\\mathcal{L}=\\mathrm{MSE}(X,\\hat{X})+\\beta\\mathcal{L}_{\\mathrm{entropy}}+\\gamma\\mathcal{L}_{\\mathrm{task}},", "mathml": "ℒ=MSE⁡(X,X^)+β​ℒentropy+γ​ℒtask,", "context": "The complete network is optimized end-to-end so that compression does not become an isolated objective that destroys information needed by the application. We therefore combine reconstruction supervision, entropy regularization, and the downstream task objective as", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E2", "label": "(2)", "section": "III Proposed Method", "role": "loss", "latex": "\\mathcal{L}_{\\mathrm{entropy}}=\\max\\left(\\mathbb{E}_{Z_{\\mathrm{quant}}\\sim P}\\left[-\\log_{2}P(Z_{\\mathrm{quant}})\\right]-H_{\\mathrm{ref}},\\,0\\right).", "mathml": "ℒentropy=max⁡(𝔼Zquant∼P​[−log2⁡P⁡(Zquant)]−Href, 0).", "context": "The entropy term acts directly on the quantized encoder latent and is defined as", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S6.E7", "label": "(7)", "section": "VI Experimental Results", "role": "metric", "latex": "R_{\\mathrm{Quant+Huff}}{=}\\frac{224\\times 224\\times 3\\times 8}{4\\times 4\\times 4\\times 1.57}\\approx 11{,}985\\times,", "mathml": "RQuant+Huff=224×224×3×84×4×4×1.57≈11,985×,", "context": "OASIS supports two complementary mechanisms for reducing off-chip communication. In the quantization-and-Huffman path, the encoder reduces the spatial and channel dimensions of the activation, after which 4-bit quantization and entropy coding further reduce the transmitted bit volume. For example, the configuration with a 4×4×4 latent representation provides", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-09-30", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }